This repository showcases some AI and ML projects implemented by Adam Smith.
This project combines deep learning and graph search to solve the Rubik’s Cube. It extends and improves the methods developed in this research. A neural network is trained to predict the distance from any scrambled cube state to the solved state, and this prediction is used as a heuristic function in a beam search algorithm.
This project performs sentiment analysis on IMDB movie ratings by classifying movie reviews as either positive or negative based on their text content. Three different models are implemented and evaluated: a Naive Bayesian Network, a Convolutional Neural Network (CNN) with pretrained GloVe embeddings, and a Fully-Connected Neural Network using Keras' built-in IMDB dataset processing.
This project analyzes images and generates artificial optical data from the PICO-60 dark matter search experiment using a generative adversarial network (GAN). Due to its low-background design, PICO-60 lacks the large quantities of data necessary for training of machine learning algorithms. This project provides a method for generating artificial optical data mimicking images obtained from the PICO-60 detector’s four cameras.
The MoonBoard is an artificial climbing wall with a fixed set of holds. Here, a Convolutional Neural Network and a Generative Adversarial Network are implemented to automatically grade climbing problems and generate new climbing routes on the MoonBoard.
This research project explores an innovative quantum machine learning approach to Land Use and Land Cover (LULC) classification using a novel quantum dropout technique applied to satellite imagery. A transfer learning approach is used to adapt ResNet-50 to classification of satellite imagery, and the novel quantum dropout layer is integrated into the fully-connected head for adaptive regularization.
This project explores the use of quantum machine learning techniques, specifically Quantum Support Vector Machines (QSVMs) and Quantum Boosting (QBoost), for the task of building detection from 3D LiDAR point cloud data. The goal is to compare the performance of these quantum models against their classical counterparts, such as Support Vector Machines (SVMs) and AdaBoost, on this spatial classification problem.
Here we implement the Hybrid Quantum-Classical Convolutional Neural Network (HQCCNN) described in this paper by Li et al. Here, we employ the HQCCNN architecture to classify land cover from the DeepSat4 dataset.
This project studies and evaluates methods for training neural networks on manifold-valued data. Current methods often neglect the manifold structure and lead to ineffective and/or inefficient networks. Preprocessing techniques and various network structures are studied for data on various manifolds with the goal of obtaining an efficient and effective neural network architecture.
This repository contains a few computer vision models developed using PyTorch and MediaPipe including face detection, face mesh, hand tracking, pose estimation, and gesture volume control models as well as a deep convolutional GAN for image deblurring.
This repository contains basic implementations of a few machine learning algorithms as well as the preliminary data analysis involved. The models include linear regression, multiple linear regression, support vector machines, and logistic regression.